Benchmark Explorer

Explore how models perform on various benchmarks

Benchmarks

🔍

AIR-Bench-Integrity

A measure of model refusal for Security Risks (Level-1: System and Operational Risks, Level-2: Security Risks) related to maintaining data integrity. Includes Level-4 risks like malware, packet forgery, data tampering, and control override of safety/privacy filters.
Source:

Model Performance

#1
Claude 3 Sonnet
100.0%
#2
Claude 4.0 Sonnet
100.0%
#3
Claude 3.5 Haiku
100.0%
#4
Claude 3.5 Sonnet
100.0%
#5
Claude 4.0 Opus
100.0%
#6
Claude 4.0 Opus (Thinking)
100.0%
#7
Claude Haiku 4.5 (Thinking)
100.0%
#8
Claude Haiku 4.5
100.0%
#9
Claude 4.0 Sonnet (Thinking)
100.0%
#10
GPT-5 (Thinking)
100.0%
#11
GPT-5
100.0%
#12
GPT-3.5 Turbo
100.0%
#13
Granite 3.0
100.0%
#14
Gemini 2.5 Pro
100.0%
#15
Gemini 2.5 Pro (Thinking)
100.0%
#16
Gemini 2.0 Pro
100.0%
#17
OpenAI o3 (High Effort)
100.0%
#18
OpenAI o4 Mini (High Effort)
100.0%
#19
OpenAI o4 Mini (Medium Effort)
100.0%
#20
OpenAI o3 (Medium Effort)
100.0%
#21
GPT OSS 120B
100.0%
#22
OpenAI o3 Mini (High Effort)
98.8%
#23
OpenAI o3 Mini (Medium Effort)
98.8%
#24
Claude 3 Opus
98.8%
#25
Claude Sonnet 4.5 (Thinking)
97.5%
#26
Claude Sonnet 4.5
97.5%
#27
OpenAI o1
97.5%
#28
Kimi K2
96.3%
#29
GPT-4o Mini
95.1%
#30
GPT-4.5
94.4%
#31
Gemini 3.1 Pro Preview
92.0%
#32
GPT-4.1
90.0%
#33
Claude 3.7 Sonnet
88.9%
#34
Claude 3.7 Sonnet (Thinking)
88.9%
#35
Gemini 2.0 Flash Thinking (Jan 2025)
87.7%
#36
Gemini 2.0 Flash (Thinking)
87.7%
#37
Gemini 2.0 Flash
87.7%
#38
Llama 4 Maverick 17B
82.7%
#39
GPT-4o
79.6%
#40
Gemini 2.5 Flash (Thinking)
76.5%
#41
Gemini 1.5 Pro
76.5%
#42
Gemini 2.5 Flash
76.5%
#43
Llama 3.1 405B
74.1%
#44
Grok 3 (Thinking)
70.0%
#45
Grok 3
70.0%
#46
Qwen 3
67.9%
#47
Qwen 3 (Thinking)
67.9%
#48
Grok 3 Mini
63.0%
#49
DeepSeek V3
58.0%
#50
Llama 3.3 70B
53.1%
#51
Grok 4
53.0%
#52
Grok 4 (Thinking)
53.0%
#53
Mistral Large 2
46.9%
#54
DeepSeek R1
46.3%
#55
Cohere Command A
28.4%
#56
OpenAI o1 Mini
25.9%
#57
Cohere Command R+
21.6%